Closing the Loop: A Self-Evolving Defense

StackOne built an autonomous loop that discovers new prompt injection techniques, scores them against real models and its Defender classifier, and files a PR back to the package to keep detection current. It follows two prior posts on training a 22MB prompt injection classifier and calibrating its false positive rate on real connector data.